# Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. # # All contributions by Andy Brock: # Copyright (c) 2019 Andy Brock # MIT License import sys import os import numpy as np import time import datetime import json import math import torch import torchvision.transforms as transforms from torch.utils.data import DataLoader import shutil import torch.distributed as dist sys.path.insert(1, os.path.join(sys.path[0], "..")) from data_utils.resnet import resnet50 import data_utils.datasets_common as dset from data_utils.cocostuff_dataset import CocoStuff class CenterCropLongEdge(object): """Crops the given PIL Image on the long edge. Parameters ---------- size: sequence or int Desired output size of the crop. If size is an int instead of sequence like (h, w), a square crop (size, size) is made. """ def __call__(self, img): """ Args: img (PIL Image): Image to be cropped. Returns: PIL Image: Cropped image. """ return transforms.functional.center_crop(img, min(img.size)) def __repr__(self): return self.__class__.__name__ # Modified to be able to do class-balancing class DistributedSampler(torch.utils.data.sampler.Sampler): """Sampler that restricts data loading to a subset of the dataset. It is especially useful in conjunction with :class:`torch.nn.parallel.DistributedDataParallel`. In such case, each process can pass a DistributedSampler instance as a DataLoader sampler, and load a subset of the original dataset that is exclusive to it. .. note:: Dataset is assumed to be of constant size. Arguments: dataset: Dataset used for sampling. num_replicas (optional): Number of processes participating in distributed training. rank (optional): Rank of the current process within num_replicas. shuffle (optional): If true (default), sampler will shuffle the indices """ def __init__( self, dataset, num_replicas=None, rank=None, shuffle=True, weights=None ): if num_replicas is None: if not torch.dist.is_available(): raise RuntimeError("Requires distributed package to be available") num_replicas = torch.dist.get_world_size() if rank is None: if not torch.dist.is_available(): raise RuntimeError("Requires distributed package to be available") rank = torch.dist.get_rank() self.dataset = dataset self.num_replicas = num_replicas self.rank = rank self.epoch = 0 self.num_samples = int(math.ceil(len(self.dataset) * 1.0 / self.num_replicas)) self.total_size = self.num_samples * self.num_replicas self.shuffle = shuffle self.weights = weights def __iter__(self): # deterministically shuffle based on epoch g = torch.Generator() g.manual_seed(self.epoch) if self.shuffle: if self.weights is not None: print("using class balanced!") indices = torch.multinomial( self.weights, len(self.dataset), replacement=True, generator=g ).tolist() else: indices = torch.randperm(len(self.dataset), generator=g).tolist() else: indices = list(range(len(self.dataset))) # add extra samples to make it evenly divisible indices += indices[: (self.total_size - len(indices))] assert len(indices) == self.total_size # subsample indices = indices[self.rank : self.total_size : self.num_replicas] assert len(indices) == self.num_samples return iter(indices) def __len__(self): return self.num_samples def set_epoch(self, epoch): self.epoch = epoch class CheckpointedSampler(torch.utils.data.Sampler): """Resumable sample with a random generated initialized with a given seed. Arguments --------- data_source: Dataset Dataset to sample from. start_itr: int, optional Number of iteration to start (or restart) the sampling. start_epoch: int, optional Number of epoch to start (or restart) the sampling. batch_size: int, optional Batch size. class_balanced: bool, optional Sample the data with a class balancing approach. custom_distrib_gen: bool, optional Use a temperature controlled class balancing. samples_per_class: list, optional A list of int values that indicate the number of samples per class. class_probabilities: list, optional A list of float values indicating the probability of a class in the dataset. longtail_temperature: float, optional Temperature value to smooth the longtail distribution with a softmax function. seed: int, optional Random seed used. """ def __init__( self, data_source, start_itr=0, start_epoch=0, batch_size=128, class_balanced=False, custom_distrib_gen=False, samples_per_class=None, class_probabilities=None, longtail_temperature=1, seed=0, ): self.data_source = data_source self.num_samples = len(self.data_source) self.start_itr = start_itr % (len(self.data_source) // batch_size) self.start_epoch = start_epoch self.batch_size = batch_size self.class_balanced = class_balanced self.custom_distrib_gen = custom_distrib_gen self.generator = torch.Generator() self.generator.manual_seed(seed) if self.class_balanced: print("Class balanced sampling.") self.weights = make_weights_for_balanced_classes( samples_per_class, self.data_source.labels, 1000, self.custom_distrib_gen, longtail_temperature, class_probabilities=class_probabilities, ) self.weights = torch.DoubleTensor(self.weights) # Resumable data loader print( "Using the generator ", self.start_epoch, " times to resume where we left off.", ) # print('Later, we will resume at iteration ', self.start_itr) for epoch in range(self.start_epoch): self._sample_epoch_perm() if not isinstance(self.num_samples, int) or self.num_samples <= 0: raise ValueError( "num_samples should be a positive integeral " "value, but got num_samples={}".format(self.num_samples) ) def _sample_epoch_perm(self): if self.class_balanced: out = [ torch.multinomial( self.weights, len(self.data_source), replacement=True, generator=self.generator, ) ] else: out = [torch.randperm(len(self.data_source), generator=self.generator)] return out def __iter__(self): out = self._sample_epoch_perm() output = torch.cat(out).tolist() return iter(output) def __len__(self): return len(self.data_source) def make_weights_for_balanced_classes( samples_per_class, labels=None, nclasses=None, custom_distrib_gen=False, longtail_temperature=1, class_probabilities=None, ): """It prepares the sampling weights for the DataLoader. Arguments --------- samples_per_class: list A list of int values (size C) that indicate the number of samples per class, for all C classes. labels: list/ NumPy array/ torch Tensor, optional A list of size N that contains a class label for each sample. nclasses: int, optional Number of classes in the dataset. custom_distrib_gen: bool, optional Use a temperature controlled class balancing. longtail_temperature: float, optional Temperature value to smooth the longtail distribution with a softmax function. class_probabilities: list A list of float values (size C) indicating the probability of a class in the dataset. seed: int Random seed used. Returns ------- If custom_distrib_gen is True, a torch Tensor with size C, where C is the number of classes, that contains the sampling weights for each class. If custom_distrib_gen is False, a list with size N (dataset size) that contains the sampling weights for each individual data sample. """ if custom_distrib_gen: # temperature controlled distribution print( "Temperature controlled distribution for balanced classes! " "Temperature:", longtail_temperature, ) class_prob = torch.log(torch.DoubleTensor(class_probabilities)) weight_per_class = torch.exp(class_prob / longtail_temperature) / torch.sum( torch.exp(class_prob / longtail_temperature) ) else: count = [0] * nclasses for item in labels: count[item] += 1 weight_per_class = [0.0] * nclasses N = float(sum(count)) for i in range(nclasses): # Standard class balancing weight_per_class[i] = N / float(count[i]) # Convert weighting per class to weighting per example weight = [0] * len(labels) for idx, val in enumerate(labels): # Uniform probability of selecting a sample, given a class # p(x|y)p(y) weight[idx] = (1 / samples_per_class[val]) * weight_per_class[val] return weight def load_pretrained_feature_extractor( pretrained_path="", feature_extractor="classification", backbone_feature_extractor="resnet50", ): """It loads a pre-trained feature extractor. Arguments --------- pretrained_path: str, optional Path to the feature extractor's weights. feature_extractor: str, optional If "classification" a network trained on ImageNet for classification will be used. If "selfsupervised", a network trained on ImageNet with self-supervision will be used. backbone_feature_extractor: str, optional Name of the backbone for the feature extractor. Currently, only ResNet50 is supported. Returns ------- A Pytorch network initialized with pre-trained weights. """ if backbone_feature_extractor == "resnet50": print("using resnet50 to extract features") net = resnet50( pretrained=False if pretrained_path != "" else True, classifier_run=False ).cuda() else: raise ValueError("Not implemented for backbones other than ResNet50.") if pretrained_path != "": print("Loading pretrained weights from: ", pretrained_path) # original saved file with DataParallel state_dict = torch.load(pretrained_path) if not feature_extractor == "selfsupervised": state_dict = state_dict["state_dict_best"]["feat_model"] # create new OrderedDict that does not contain `module.` from collections import OrderedDict new_state_dict = OrderedDict() for k, v in state_dict.items(): if "module." in k: name = k[7:] # remove `module.` elif "_feature_blocks." in k: name = k.replace("_feature_blocks.", "") else: name = k if name in net.state_dict().keys(): new_state_dict[name] = v else: print("key ", name, " not in dict") for key in net.state_dict().keys(): if key not in new_state_dict.keys(): print("Network key ", key, " not in dict to load") if not feature_extractor == "selfsupervised": state_dict = torch.load(pretrained_path)["state_dict_best"]["classifier"] # create new OrderedDict that does not contain `module.` for k, v in state_dict.items(): name = k[7:] # remove `module.` new_state_dict[name] = v # load params net.load_state_dict( new_state_dict, strict=False if feature_extractor == "selfsupervised" else True, ) else: print("Using pretrained weights on full ImageNet.") return net def get_dataset_images( resolution, data_path, load_in_mem=False, augment=False, longtail=False, split="train", test_part=False, which_dataset="imagenet", instance_json="", stuff_json="", **kwargs ): """It prepares a dataset that reads the files from a folder. Arguments --------- resolution: int Dataset resolution. data_path: str Path where to find the data. load_in_mem: bool, optional If True, load all data in memory. augment: bool, optional If True, use horizontal flips as data augmentation. longtail: bool, optional If True, use the longtailed version of ImageNet (ImageNet-LT). split: str, optional Split name to use. test_part: bool, optional Only used for COCO-Stuff. If True, use the evaluation set instead of the validation set. which_dataset: str, optional Dataset name. instance_json: str, optional Path where to find the JSON data for COCO-Stuff instances. stuff_json: str, optional Path where to find the JSON data for COCO-Stuff stuff. Returns ------- A Dataset class. """ # Data transforms norm_mean = [0.5, 0.5, 0.5] norm_std = [0.5, 0.5, 0.5] if which_dataset not in ["coco"]: transform_list = [CenterCropLongEdge(), transforms.Resize(resolution)] else: transform_list = [transforms.Resize(resolution)] transform_list = transforms.Compose( transform_list + [transforms.ToTensor(), transforms.Normalize(norm_mean, norm_std)] ) if augment: transform_list = transforms.Compose( transform_list + [transforms.RandomHorizontalFlip()] ) if which_dataset not in ["coco"]: which_dataset_file = dset.ImageFolder dataset_kwargs = {} else: print("Using coco-stuff dataset class") which_dataset_file = CocoStuff dataset_kwargs = { "image_dir": data_path, "instances_json": instance_json, "stuff_json": stuff_json, "image_size": resolution, "iscrowd": True if split == "train" else False, "test_part": test_part, } dataset = which_dataset_file( root=data_path, transform=transform_list, load_in_mem=load_in_mem, split=split, longtail=longtail, **dataset_kwargs ) return dataset def get_dataset_hdf5( resolution, data_path, augment=False, longtail=False, local_rank=0, copy_locally=False, ddp=True, tmp_dir="", class_cond=True, instance_cond=False, feature_extractor="classification", backbone_feature_extractor="resnext50", which_nn_balance="instance_balance", which_dataset="imagenet", split="train", test_part=False, kmeans_subsampled=-1, n_subsampled_data=-1, feature_augmentation=False, filter_hd=-1, k_nn=50, load_in_mem_feats=False, compute_nns=False, **kwargs ): """It prepares a dataset that reads the data from HDF5 files. Arguments --------- resolution: int Dataset resolution. data_path: str Path where to find the data. load_in_mem: bool, optional If True, load all data in memory. augment: bool, optional If True, use horizontal flips as data augmentation. longtail: bool, optional If True, use the longtailed version of ImageNet (ImageNet-LT). local_rank: int, optional Index indicating the rank of the DistributedDataParallel (DDP) process in the local machine. It is set to 0 by default or if DDP is not used. copy_locally: bool, optional If true, the HDF5 files will be copied locally to the machine. Useful if the data is in a server. ddp: bool, optional If True, use DistributedDataParallel (DDP). tmp_dir: str, optional Path where to copy the dataset HDF5 files locally. class_cond: bool, optional If True, the dataset will load the labels of the neighbor real samples. instance_cond: bool, optional If True, the dataset will load the instance features. feature_extractor: str, optional If "classification" a network trained on ImageNet for classification will be used. If "selfsupervised", a network trained on ImageNet with self-supervision will be used. backbone_feature_extractor: str, optional Name of the backbone for the feature extractor. Currently, only ResNet50 is supported. which_nn_balance: str, optional Whether to sample an instance or a neighbor class first. By default, ``instance_balance`` is used. Using ``nnclass_balance`` allows class balancing to be applied. split: str, optional Split name to use. test_part: bool, optional Only used for COCO-Stuff. If True, use the evaluation set instead of the validation set. kmeans_subsampled: int, optional If other than -1, that number of data points are selected with k-means from the dataset. It reduces the amount of available data to train or test the model. n_subsampled_data: int, optional If other than -1, that number of data points are randomly selected from the dataset. It reduces the amount of available data to train or test the model. feature_augmentation: bool, optional Use the instance features of the flipped ground-truth image instances as conditioning, with a 50% probability. filter_hd: int, optional Only used for COCO-Stuff dataset. If -1, all COCO-Stuff evaluation set is used. If 0, only images with seen class combinations are used. If 1, only images with unseen class combinations are used. k_nn: int, optional Size of the neighborhood obtained with the k-NN algorithm. load_in_mem_feats: bool, optional Load all instance features in memory. compute_nns: bool, optional If True, compute the nearest neighbors. If False, load them from a file with pre-computed neighbors. Returns ------- A Dataset class. """ if which_dataset in ["imagenet", "imagenet_lt"]: dataset_name_prefix = "ILSVRC" elif which_dataset == "coco": dataset_name_prefix = "COCO" else: dataset_name_prefix = which_dataset # HDF5 file name hdf5_filename = "%s%i%s%s%s" % ( dataset_name_prefix, resolution, "" if not longtail else "longtail", "_val" if split == "val" else "", "_test" if test_part else "", ) # Data paths data_path_xy = os.path.join(data_path, hdf5_filename + "_xy.hdf5") data_path_feats, data_path_nns, kmeans_file = None, None, None if instance_cond: data_path_feats = os.path.join( data_path, hdf5_filename + "_feats_%s_%s.hdf5" % (feature_extractor, backbone_feature_extractor), ) if not compute_nns: data_path_nns = os.path.join( data_path, hdf5_filename + "_feats_%s_%s_nn_k%i.hdf5" % (feature_extractor, backbone_feature_extractor, k_nn), ) # Load a file with indexes of the samples selected with k-means. if kmeans_subsampled > -1: if which_dataset == "imagenet": d_name = "IN" elif which_dataset == "coco": d_name = "COCO" else: d_name = which_dataset kmeans_file = ( d_name + "_res" + str(resolution) + "_rn50_" + feature_extractor + "_kmeans_k" + str(kmeans_subsampled) + ".npy" ) kmeans_file = os.path.join(data_path, kmeans_file) # Optionally copy the data locally in the cluster. if copy_locally: tmp_file = os.path.join(tmp_dir, hdf5_filename + "_xy.hdf5") print(tmp_file) if instance_cond: tmp_file_feats = os.path.join( tmp_dir, hdf5_filename + "_feats_%s_%s.hdf5" % (feature_extractor, backbone_feature_extractor), ) print(tmp_file_feats) # Only copy locally for the first device in each machine if local_rank == 0: # device == 'cuda:0': shutil.copy2(data_path_xy, tmp_file) if instance_cond: shutil.copy2(data_path_feats, tmp_file_feats) data_path_xy = tmp_file if instance_cond: data_path_feats = tmp_file_feats # Wait for the main process to copy the data locally if ddp: dist.barrier() # Data transforms if augment: transform_list = transforms.RandomHorizontalFlip() else: transform_list = None dataset = dset.ILSVRC_HDF5_feats( root=data_path_xy, root_feats=data_path_feats, root_nns=data_path_nns, transform=transform_list, load_labels=class_cond, load_features=instance_cond, load_in_mem_images=False, load_in_mem_labels=True, load_in_mem_feats=load_in_mem_feats, k_nn=k_nn, which_nn_balance=which_nn_balance, kmeans_file=kmeans_file, n_subsampled_data=n_subsampled_data, feature_augmentation=feature_augmentation, filter_hd=filter_hd, ) return dataset def filter_by_hd(ood_distance): """Pre-select image indexes in COCO-Stuff evaluation set according to its class composition. Parameters ---------- ood_distance: int Minimum hamming distance (HD) between the set of classes present in the evaluation image and all training images. If 0, pre-selected images will be the ones that only contain class sets already seen during training. If other than 0, all other images with unseen class sets will be selected, regardless of the hamming distance (HD>0). Returns ------- List of pre-selected images. """ image_ids_original = np.load( "../coco_stuff_val_indexes/cocostuff_val2_all_idxs.npy", allow_pickle=True ) print("Filtering new ids!") odd_image_ids = np.load( os.path.join( "../coco_stuff_val_indexes", "val2" + "_image_ids_by_hd_75ktraining_im.npy" ), allow_pickle=True, ) if ood_distance == 0: image_ids = odd_image_ids[ood_distance] else: total_img_ids = [] for ood_dist in range(1, len(odd_image_ids)): total_img_ids += odd_image_ids[ood_dist] image_ids = total_img_ids allowed_idxs = [] for i_idx, id in enumerate(image_ids_original): if id in image_ids: allowed_idxs.append(i_idx) allowed_idxs = np.array(allowed_idxs) print("Num images after filtering ", len(allowed_idxs)) return allowed_idxs def get_dataloader( dataset, batch_size=64, num_workers=8, shuffle=True, pin_memory=True, drop_last=True, start_itr=0, start_epoch=0, use_checkpointable_sampler=False, use_balanced_sampler=False, custom_distrib_gen=False, samples_per_class=None, class_probabilities=None, seed=0, longtail_temperature=1, rank=0, world_size=-1, **kwargs ): """Get DataLoader to iterate over the dataset. Parameters ---------- dataset: Dataset Class with the specified dataset characteristics. batch_size: int, optional Batch size. num_workers: int, optional Number of workers for the dataloader. shuffle: bool, optional If True, the data is shuffled. If a sampler is used (use_checkpointable_sampler=True, use_balanced_sampler=True or world_size>-1), this parameter is not used. pin_memory: bool, optional Pin memory in the dataloader. drop_last: bool, optional Drop last incomplete batch in the dataloader. start_itr: int, optional Iteration number to resume the sample from. Only used with use_checkpointable_sampler=True. start_epoch: int, optional Epoch number to resume the sample from. Only used with use_checkpointable_sampler=True. use_checkpointable_sampler: bool, optional If True, use the CheckpointedSampler class to resume jobs from the last seen batch (deterministic). use_balanced_sampler: bool, optional If True, balance the data according to a specific class distribution. Use in conjunction with ``custom_distrib_gen``, ``samples_per_class``, ``class_probabilities`` and ``longtail_temperature``. custom_distrib_gen: bool, optional Use a temperature controlled class balancing. samples_per_class: list, optional A list of int values that indicate the number of samples per class. class_probabilities: list, optional A list of float values indicating the probability of a class in the dataset. longtail_temperature: float, optional Temperature value to smooth the longtail distribution with a softmax function. seed: int, optional Random seed used. rank: int, optional Rank of the current process (if using DistributedDataParallel training). world_size: int, optional World size (if using DistributedDataParallel training). Returns ------- An instance of DataLoader. """ # Prepare loader; the loaders list is for forward compatibility with # using validation / test splits. # if use_multiepoch_sampler: loader_kwargs = { "num_workers": num_workers, "pin_memory": pin_memory, "drop_last": drop_last, } print("Dropping last batch? ", drop_last) # Otherwise, it has issues dividing the batch for accumulations # if longtail: # loader_kwargs.update({'drop_last': drop_last}) if use_checkpointable_sampler: print( "Using checkpointable sampler from start_itr %d..., using seed %d" % (start_itr, seed) ) sampler = CheckpointedSampler( dataset, start_itr, start_epoch, batch_size, class_balanced=use_balanced_sampler, custom_distrib_gen=custom_distrib_gen, longtail_temperature=longtail_temperature, samples_per_class=samples_per_class, class_probabilities=class_probabilities, seed=seed, ) loader = DataLoader( dataset, batch_size=batch_size, sampler=sampler, shuffle=False, worker_init_fn=seed_worker, **loader_kwargs ) else: if use_balanced_sampler: print("Balancing real data! Custom? ", custom_distrib_gen) weights = make_weights_for_balanced_classes( samples_per_class, dataset.labels, 1000, custom_distrib_gen, longtail_temperature, class_probabilities=class_probabilities, ) weights = torch.DoubleTensor(weights) else: weights = None if world_size == -1: if use_balanced_sampler: sampler = torch.utils.data.sampler.WeightedRandomSampler( weights, len(weights) ) shuffle = False else: sampler = None else: sampler = DistributedSampler( dataset, num_replicas=world_size, rank=rank, weights=weights ) shuffle = False print("Loader workers?", loader_kwargs, " with shuffle?", shuffle) loader = DataLoader( dataset, batch_size=batch_size, shuffle=shuffle, sampler=sampler, worker_init_fn=seed_worker if use_checkpointable_sampler else None, **loader_kwargs ) return loader def sample_conditioning_values( z_, y_, ddp=False, batch_size=1, weights_sampling=None, dataset=None, constant_conditioning=False, class_cond=True, instance_cond=False, nn_sampling_strategy="instance_balance", ): """It samples conditionings from the noise distribution and dataset statistics. Arguments --------- z_: Distribution Noise distribution. y_: Distribution Labels distribution ( ddp: bool, optional If True, use DistributedDataParallel (DDP). batch_size: int, optional Batch size. weights_sampling: NumPy array, optional Weights to balance the sampling of the conditionings. dataset: Dataset Instance of a dataset. constant_conditioning: bool, optional If True, set all labels to zero. class_cond: bool, optional If True, the dataset will load the labels of the neighbor real samples. instance_cond: bool, optional If True, the dataset will load the instance features. nn_sampling_strategy: str, optional Whether to sample an instance or a neighbor class first. By default, ``instance_balance`` is used. Using ``nnclass_balance`` allows class balancing to be applied. Returns ------- If not using labels (class_cond=False) nor instance features (instance_cond=False), return the sampled noise vectors. If not using labels (class_cond=False), return the sampled noise vectors and instance feature vectors, sampled according to the ``nn_sampling_strategy`` and ``weights_sampling``. If using labels (class_cond=True), return the sampled noise vectors, instance feature vectors and the neighbor class labels. """ with torch.no_grad(): z_.sample_() if not class_cond and not instance_cond: return z_ elif class_cond and not instance_cond: y_.sample_() if constant_conditioning: return z_, torch.zeros_like(y_) else: if ddp: return z_, y_ else: return z_, y_.data.clone() else: if nn_sampling_strategy == "instance_balance": sampling_funct_name = dataset.sample_conditioning_instance_balance elif nn_sampling_strategy == "nnclass_balance": sampling_funct_name = dataset.sample_conditioning_nnclass_balance labels_g, f_g = sampling_funct_name(batch_size, weights_sampling) if instance_cond and not class_cond: return z_, f_g elif instance_cond and class_cond: return z_, labels_g, f_g # Convenience function to prepare a z and y vector def prepare_z_y( G_batch_size, dim_z, nclasses, device="cuda", fp16=False, z_var=1.0, longtail_gen=False, custom_distrib=False, longtail_temperature=1, class_probabilities=None, ): """Prepare the noise and label distributions. Arguments --------- G_batch_size: int Batch size for the generator. dim_z: int Noise vector dimensionality. nclasses: int Number of classes in the dataset fp16: bool, optional Float16. z_var: float, optional Variance for the noise normal distribution. longtail_gen: bool, optional If true, use the longtail distribution for the classes (ImageNet-LT) custom_distrib: bool, optional If true, use a temperature annealed class distribution. longtail_temperature: float, optional Temperature value to smooth the longtail distribution with a softmax function. class_probabilities: list, optional A list of float values indicating the probability of a class in the dataset. Returns ------- The noise and class distributions. """ z_ = Distribution(torch.randn(G_batch_size, dim_z, requires_grad=False)) z_.init_distribution("normal", mean=0, var=z_var) # z_ = z_.to(device, torch.float16 if fp16 else torch.float32) if fp16: z_ = z_.half() y_ = Distribution(torch.zeros(G_batch_size, requires_grad=False)) if longtail_gen: y_.init_distribution( "categorical_longtail", num_categories=nclasses, class_prob=class_probabilities, ) elif custom_distrib: y_.init_distribution( "categorical_longtail_temperature", num_categories=nclasses, temperature=longtail_temperature, class_prob=class_probabilities, ) else: y_.init_distribution("categorical", num_categories=nclasses) # y_ = y_.to(device, torch.int64) return z_, y_ # A highly simplified convenience class for sampling from distributions # One could also use PyTorch's inbuilt distributions package. # Note that this class requires initialization to proceed as # x = Distribution(torch.randn(size)) # x.init_distribution(dist_type, **dist_kwargs) # x = x.to(device,dtype) # This is partially based on https://discuss.pytorch.org/t/subclassing-torch-tensor/23754/2 class Distribution(torch.Tensor): # Init the params of the distribution def init_distribution(self, dist_type, class_prob=None, **kwargs): self.dist_type = dist_type self.dist_kwargs = kwargs if self.dist_type == "normal": self.mean, self.var = kwargs["mean"], kwargs["var"] elif self.dist_type == "categorical": self.num_categories = kwargs["num_categories"] elif self.dist_type == "categorical_longtail": print("(class conditioning sampler) using longtail distribution") self.num_categories = kwargs["num_categories"] self.class_prob = torch.DoubleTensor(class_prob) elif self.dist_type == "categorical_longtail_temperature": print( "(class conditioning sampler) Softening the long-tail distribution with temperature ", kwargs["temperature"], ) self.num_categories = kwargs["num_categories"] self.class_prob = torch.log(torch.DoubleTensor(class_prob)) self.class_prob = torch.exp( self.class_prob / kwargs["temperature"] ) / torch.sum(torch.exp(self.class_prob / kwargs["temperature"])) def seed_generator(self, seed): self.generator.manual_seed(seed) def sample_(self): if self.dist_type == "normal": self.normal_(self.mean, self.var) elif self.dist_type == "categorical": self.random_(0, self.num_categories) elif ( "categorical_longtail" in self.dist_type or "categorical_longtail_temperature" in self.dist_type ): self.data = torch.multinomial( self.class_prob, len(self), replacement=True ).to(self.device) # return self.variable # Silly hack: overwrite the to() method to wrap the new object # in a distribution as well # def to(self, *args, **kwargs): # new_obj = Distribution(self) # new_obj.init_distribution(self.dist_type, **self.dist_kwargs) # new_obj.data = super().to(*args, **kwargs) # return new_obj def seed_worker(worker_id): worker_seed = torch.initial_seed() + worker_id